Speakers encode silent structures: Evidence from complementizer priming in English

Do speakers encode abstract structural representations devoid of perceptual-motor content, that is, phonology? In six recall-based production experiments, we examined whether English speakers encode the null complementizer in sentence production using structural priming, the tendency for speakers to reuse the structure they have recently encountered. The results show that the null complementizer can be primed across distinct construction types and that this priming effect cannot be explained as the priming of the absence of the overt complementizer. These results are difficult to capture in semantic, pragmatic, or phonological terms. Furthermore, we evaluated two varieties of neural network language models (based on transformers and long short term memory) for their capacity to reproduce human priming patterns. Although they could reproduce basic priming effects, neural network language models were simultaneously more sensitive to constructional differences and less sensitive to abstract similarities across constructions than humans. This suggests that distributional cues alone are likely not sufficient for learning the generalization governing the distribution of English complementizers. Based on these results, we argue that the structural representations speakers construct during production go beyond what they hear and say.